Edge Machine Learning for Face Detection

Abstract

This paper describes an implementation of edge machine learning for vision-based classification and detection tasks. In edge machine learning, machine and deep learning algorithms are executed locally on embedded devices rather than on more powerful computers or the cloud. The main task explored is face detection using a low-power microcontroller. This device utilizes a convolutional neural network (CNN) accelerator that optimizes convolution and pooling operations for fast power-efficient inference. Development for this system requires building and training a hardwarelimited CNN rather than fine-tuning a pre-trained state-of-the-art model. The development process is discussed along with the constraints of this embedded device.